Incorporating AI into your Workflow

Think, Pair, Share

For the following use cases, discuss whether you think AI is appropriate to use as part of your research:

• Developing research questions/ideas
• Coding
• Theoretical modeling and derivations
• Editing
• Writing
• Summarizing (post-publication)

Two observations and a question

A few thoughts

• The average quality (and expectations) of a JMP will be higher every year
• Expectations for more complex analyses will increase
• Bad writing will likely become a very bad signal
• Boring/generic writing will signal overreliance on AI
• Interpersonal skills will become more critical

Finding the Balance

1. Learn the fundamentals — You can’t evaluate code you don’t understand. AI will confidently give you wrong answers—you need to catch them. Think of AI as a fast RA who occasionally hallucinates.

2. But don’t fall behind — Your peers are using these tools. The productivity gap compounds quickly. “I prefer the old way” is not a competitive strategy.

3. The old system was imperfect — Research errors happened before AI—and were caught late or not at all. Better workflows + AI review can catch errors earlier.

The old system wasn’t a gold standard

Case 1: Reinhart & Rogoff (2010) — An Excel formula didn’t include 5 rows of data. Paper claimed high debt causes -0.1% growth; corrected: +2.2% growth. Used to justify global austerity. Found by a grad student, 3 years later.

Case 2: Deschênes & Greenstone (2007 AER) — Coding errors and missing weather data in climate-agriculture analysis. Paper claimed climate change would increase ag profits; corrected results showed the opposite. Found 5 years later via replication.

1. Use a cloud platform + clear folder structure

Work from OneDrive, Dropbox, or Google Drive—not your local machine alone. A consistent folder structure makes AI tools more effective because they can understand your project organization. This also enables collaboration and automatic backup simultaneously.

project-folder/
├── data/
│   ├── raw/
│   └── clean/
├── code/
│   ├── 01-clean.R
│   ├── 02-analysis.R
│   └── 03-figures.R
├── output/
├── paper/
└── README.md

2. Use Git/GitHub for version control

2. Use Git/GitHub for version control

No more analysis_v3_final_FINAL2.R — Git tracks it all.

git diff analysis.R

- model <- lm(y ~ x1 + x2, data = df)
+ model <- lm(y ~ x1 + x2 + x1:x2, data = df)  # added interaction

Confused on how to set up Git and GitHub? Ask AI!

3. Use AI coding tools: Claude Code vs. Cursor

4. Leveling Up: Structured Workflows

Basic Claude Code is powerful, but can be unpredictable.

Structured workflows add:

• CLAUDE.md — Project config Claude reads every session
• Plan-first approach — Review the approach before implementation
• Quality gates — Standards enforced before committing

The .claude/ Folder Structure

my-project/
├── CLAUDE.md                    # Project "constitution"
├── .claude/
│   └── rules/
│       ├── r-code-conventions.md    # R coding standards
│       ├── quality-gates.md         # 80/90/95 scoring
│       └── plan-first-workflow.md   # Plan → Approve → Implement
├── code/
├── data/
│   ├── raw/
│   └── clean/
└── output/

CLAUDE.md is read at the start of every session — Claude knows your conventions automatically.

The Plan-First Workflow

The pattern:

1. Describe — What you want to accomplish
2. Claude plans — Proposes approach, files, packages
3. You review — Approve or refine
4. Claude implements — Works autonomously after approval
5. Verify & commit — Quality gates checked before commit

Quality Gates

Every file gets a score (0-100). Scores below threshold block the action:

5. Write papers in Overleaf + GitHub

• Overleaf = collaborative LaTeX editing in the browser
• GitHub sync = version control for your paper
• AI tools can help draft, edit, and format LaTeX
• Figures from code → paper in one pipeline

The workflow:

Code (R/Python) → Figures & Tables → GitHub Repo → Overleaf Paper

One integrated system: Code changes automatically flow through to your paper.

Cautions & Pitfalls

1. AI can hallucinate packages and functions — It will confidently suggest code using libraries that don’t exist. Always run and verify.

2. Don’t let AI do your economic thinking — It’s a tool, not a co-author. Identification, intuition, and interpretation are yours.

3. Always test AI-generated code — Run it. Check edge cases. Verify results make sense.

4. Be thoughtful about sensitive data — Know what data you’re sending to AI services. Check your IRB and data use agreements.

5. Don’t skip the planning step — Review Claude’s plan before implementation. Easier to fix a wrong plan than wrong code.

More Resources

Questions?